DP
RIETI Discussion Paper Series 12-E-022
The Effects of a Megabank Merger on Firm-Bank Relationships
and Borrowing Costs
UCHINO Taisuke
RIETI
UESUGI Iichiro
RIETI
The Research Institute of Economy, Trade and Industry
http://www.rieti.go.jp/en/
RIETI Discussion Paper Series 12-E-022
April 2012
The Effects of a Megabank Merger on Firm-Bank Relationships
and Borrowing Costs *
UCHINO Taisuke †
Research Institute of Economy, Trade and Industry／Daito Bunka University
UESUGI Iichiro ‡
Research Institute of Economy, Trade and Industry／Hitotsubashi University
Abstract
Using a unique dataset of non-listed firms that identifies the banks with which firms transact, we examine
the effects of the largest-ever bank merger in Japan—that between Bank of Tokyo-Mitsubishi (BTM) and UFJ
Bank (UFJ) in 2005. We focus on how the merger affected firms through their firm-bank relationships.
Specifically, we examine whether there are any differences in how the availability of loans evolved over time
for firms that, prior to the merger, either transacted with both of the merged banks, with one of them, or with
none. We find the following: (1) Firms that had transacted with both BTM and UFJ saw their borrowing costs
increase by 40bp relative to those that had transacted with neither. (2) Firms that transacted with one of the two
banks saw their borrowing costs increase by a smaller but still significant margin of 20bp relative to those that
had transacted with neither. And (3) we do not find a significant difference in the extent that borrowing costs
increased between firms that transacted with the acquiring bank (BTM) and those that transacted with the
acquired bank (UFJ). These results suggest that the bank merger increased firms’ borrowing costs partly
through an exogenous decrease in the number of firm-bank relationships and partly through changes in the
organizational structure of the merged bank, including a consolidation of the branch network.
Keywords: Firm-bank relationship; Interest rates; Switching costs.
JEL classification: G21, G34
RIETI Discussion Papers Series aims at widely disseminating research results in the form of professional
papers, thereby stimulating lively discussion. The views expressed in the papers are solely those of the
author(s), and do not represent those of the Research Institute of Economy, Trade and Industry.
*
We would like to thank Makoto Saito, Etsuro Shioji, Kozo Harimaya, Ralph Paprzycki, and participants of the
Macro Lunch Workshop and the Labor and Industry Workshop at Hitotsubashi University, the JEA 2010
Autumn Meeting at Kwansei Gakuin University, the Regional Finance Conference at Kobe University, the
Finance Workshop at Ritsumeikan University, and the RIETI Study Group on Financial Networks for helpful
comments.
We would like to thank RIETI for the use of the dataset. This research was supported by the Global COE
program of Hitotsubashi University.
† Research Institute of Economy, Trade and Industry. Email: uchino-taisuke@rieti.go.jp.
‡ Institute of Economic Research, Hitotsubashi University, and Research Institute of Economy, Trade and
Industry. Email: iuesugi@ier.hit-u.ac.jp.
1
Introduction
Japan, like other developed countries, has seen a wave of bank mergers in the past 20 years.
Mainly as a result of these mergers, the number of banks has decreased substantially over this
period, falling from 1,069 in 1990 to 595 in 2010. Especially the number of so-called city banks banks that operate nation-wide and across national borders has declined signi…cantly, dropping
from 13 in 1990 to only 4 in 2010.
Policymakers have been concerned about the implications of these bank mergers. Since bank
consolidation changes the market structure for loans and …rm-bank relationships, a particular
focus of such concerns is whether e¢ ciency gains from mergers are passed on to borrowers or
are appropriated by the merged banks. Studies seeking to address these issues empirically, such
as Sapienza (2002) and Erel (2011), have typically focused on the availability and the price of
loans provided by the merged banks. The present study follows this approach, using micro-level
data to investigate how the availability of loans and the interest rates that …rms have to pay
develop after a major merger.
For our analysis, we use the merger of two major city banks, Bank of Tokyo-Mitsubishi
(BTM) and UFJ Bank (UFJ), in 2005. The merger of these banks, which took place toward the
end of the bank merger wave in Japan, had a sizable impact on the domestic loan market. The
combined amount of loans outstanding of the two banks in 2004 stood at 70 trillion yen (700
billion US dollar at the exchange rate at the time), equivalent to 18% of total loans outstanding
extended by …nancial institutions in Japan, making the newly-formed entity, the Bank of TokyoMitsubishi UFJ (BTMU), the largest …nancial institution in the country.
As pointed out by Williamson (1968) in his theoretical analysis, horizontal mergers in general
are likely to increase market power, raise prices for goods and services, and decrease the surplus
of …rms or consumers demanding the goods and services in the market. This expected negative
e¤ect of mergers has been the focus of many previous studies on bank mergers. Conducting
a comprehensive review of over 250 studies, Berger, Demsetz, and Strahan (1999) …nd that
the empirical evidence suggests that some types of consolidation indeed result in an increase in
market power.
2
However, in the market for bank loans, the interactions among market power, market structure, and transaction conditions are complicated by the fact that transactions between banks
and borrowers often are not at arm’s length but involve some sort of relationship. Suppose,
for example, that information on the viability of borrower …rms is produced through bank-…rm
relationships and cannot be transferred to potential new lenders. In this case, it is unlikely
that all lenders, both incumbents and potential new lenders, can exercise market power and set
loan conditions in a similar manner. Instead, bank-…rm relationships or, in Sapienza’s (2002)
terminology, “information-based market power” come to play an important role in determining
loan conditions, including interest rates.
Against this background, the primary focus of our study is to examine how a bank merger
a¤ects …rm-bank relationships and changes the availability of loans after the merger. By focusing
on the merger of two megabanks, each with a large number of borrower …rms, we are able to
classify …rms into three categories: …rms that transacted with both banks at the time of the
merger (BTM and UFJ in the case we are examining); those that transacted with one of the two
banks; and …rms that transacted with neither of these banks. Classifying …rms in this manner,
we examine how a bank merger a¤ects the availability of loans through …rm-bank relationships.
Speci…cally, we focus on the following issues. First, for …rms in the …rst category, the merger
automatically reduces their number of …rm-bank relationships by one, which is likely to have an
impact on the loan conditions they face. Studies such as those by Petersen and Rajan (1994)
and Harho¤ and Körting (1998) regard the number of …rm-bank relationships as a proxy for the
switching costs a …rm faces (the larger the number of relationships, the smaller the switching
costs) and examine the link between the number of relationships and the availability of credit.
The BTM-UFJ merger provides an excellent opportunity to examine the impact of an exogenous
reduction in the number of relationships on loan availability and borrowing costs.
Second, the bank merger is likely to have had a greater impact on …rms in the …rst two
categories (…rms that transacted with both or one of the merged banks) than those in the
third category (…rms that transacted with neither) due to changes in organizational structure
and the reallocation of resources to increase managerial e¢ ciency at the newly merged bank.
Potential organizational changes as a result of the bank merger include the consolidation of
3
branches, directly a¤ecting …rms that had a relationship with one or both of the merged banks
and maintained a relationship with the merged entity. In the case of the BTM- UFJ merger,
an extensive consolidation of the branch network followed, providing an excellent case study to
examine this issue.
Third, within the second category of …rms, there may well be a di¤erence in the way loans
are provided between borrower …rms that transacted with the acquiring bank and those that
transacted with the acquired bank. As pointed out by Peek and Rosengren (1998), the lending
behaviour of a newly merged bank usually mirrors that of the acquiring bank. Hence, …rms
that transacted with the acquired bank may face more stringent loan conditions following the
merger than those that transacted with the acquiring bank. Indirect evidence of such di¤erences
in loan conditions depending on whether a …rm transacted with the acquiring or acquired bank
is provided by Karceski, Ongena, and Smith (2005), who used share price data of listed …rms
in Norway. Here, we examine the same issue using our database. Since the …rms that we focus
on, non-listed …rms, are weaker in loan negotiations than the listed …rms examined by Karceski,
Ongena, and Smith (2005), one would expect the impact to be more pronounced, and we use
more direct measures than the share prices to examine the impact on loan conditions.
Further, we consider the results of these examinations in the context of the literature on the
“ine¢ cient” Japanese banking sector in the late 1990s and early 2000s. Peek and Rosengren
(2005) and Caballero, Hoshi, and Kashyap (2008), for example, reported that …nancial institutions had perverse incentives to postpone the realization of loan-losses caused by non-performing
loans. The measures adopted to this end by …nancial institutions include forbearance lending
and accepting reduced loan repayments. As a result, as shown by Smith (2003), the interest
rate charged by Japanese …nancial institutions did not necessarily re‡ect the credit risk of a
borrower …rm. Under these circumstances, it is worth examining if the bank merger not only
changed the level of borrowing costs but also changed the risk-interest rate relationship in the
Japanese loan market.
For the analysis, we employ a unique …rm-level panel dataset for the years 2004-2008 of
about 11,000 non-listed …rms that mostly depend on bank loans for external …nance. The
dataset not only contains balance-sheet information, which we use for generating our variables
4
on loan availability and …rm performance, but also provides the names of up to ten banks and
their branches that a …rm transacts with and that we use for identifying …rms’ relationship
with BTM and UFJ. To measure the e¤ect of the merger of the two banks on …rms, we employ
the propensity score matching di¤erence-in-di¤erences estimator (PSM-DID). Since there are
multiple ways in which the bank merger a¤ects …rms’ con…guration of bank relationships, we
follow the procedure proposed by Lechner (2002) and allow for multiple treatments.
Our …ndings can be summarized as follows. First, …rms that transacted with both BTM
and UFJ prior to the merger saw their borrowing costs increase by 40 basis points (bp) relative
to …rms that had no such transaction relationships. Second, …rms that transacted with one of
the two banks saw their borrowing costs increase by a smaller but still signi…cant margin of
20bp. Third, we do not …nd a signi…cant di¤erence in the extent that borrowing costs increased
between …rms that transacted with the acquiring bank (BTM) and those that transacted with
the acquired bank (UFJ). Overall, these results indicate that the bank merger increased …rms’
borrowing costs partly through the exogenous decrease in the number of …rm-bank relationships
and partly through changes in organizational structure at the merged bank. Finally, the treatment e¤ect does not signi…cantly di¤er depending on …rms’riskiness, meaning that our results
provide no evidence that the interest rate that a …rm is charged better re‡ects …rms’credit risk
than before the merger.
The remainder of the paper is organized as follows. Section 2 provides an overview of previous
studies and presents the empirical hypotheses. Section 3 then presents details on the merger
of BTM and UFJ and its background. Next, Sections 4 and 5 respectively outline the dataset
used for the analysis and the empirical approach. Section 6 provides summary statistics and the
estimation results followed by a further discussion in Section 7. Section 8 concludes.
2
Previous Studies and Empirical Hypotheses
There are a considerable number of studies examining the impact of bank mergers on …rms’
borrowing conditions, such as Sapienza (2002) and Erel (2011), who focus on the increasing
market concentration in the loan market following a merger. An increase in market concentration
through a merger not only results in an increase in market power of the merged banks (through
5
the increase in the market share of loans outstanding) but also of all other banks operating in
the market, allowing them to impose more stringent loan conditions and extracting larger rent
than before the merger.
At the same time, however, the interactions among market structure, the degree of competition, and transaction conditions are not straightforward due to the role of bank-borrower
relationships. The literature on …nancial intermediation suggests that banks produce valuable
information on the viability of borrowers through bank-borrower relationships (Diamond 1984).
If such information is not transferable to potential new lenders, borrowers will incur switching
costs when terminating a relationship with an existing lender and …nding a new lender (Rajan
1992; Sharpe 1990). Firm-bank relationships can be found in most advanced economies, although their duration and importance tends to vary. Reviewing a number of empirical studies
on a range of countries, Degryse, Kim, and Ongena (2008) suggest that the average duration of
…rm-bank relationships ranges from 8 years in Belgium (Degryse and Van Cayseele 2000) and
7 to 11 years in the United States (Petersen and Rajan 1994) to 13 years in Germany (Harho¤
and Körting 1998), 14 years in Italy (Angelini, Di Salvo, and Ferri 1998), and 15 to 18 years
in Norway (Ongena and Smith 2001). Meanwhile, studies on Japan suggest that the average
duration is over 20 years (e.g., Horiuchi, Packer, and Fukuda 1988; Uchida, Udell, and Watanabe
2006), suggesting that particularly in Japan, …rm-bank relationships appear to play a key role
in the loan market and that, consequently, it is important to focus on such relationships when
examining the terms and conditions of loan contracts in Japan.
Changes in …rm-bank relationships following a merger have three characteristics which can be
directly linked to empirical hypotheses. First, a bank merger results in an exogenous decrease in
the number of …rm-bank relationships. For …rms that used to transact with both of the merging
banks, the number of banks they transact with automatically declines by one after the merger.
In contrast, the number of unique borrower …rms for the merged banks remains the same after
the merger.
Regarding the e¤ect of the number of relationships, many studies, such as those by Petersen
and Rajan (1994) and Harho¤ and Körting (1998), examine the hypothesis that a smaller number of …rm-bank relationships means that switching costs for borrowers are higher, endowing
6
banks with “information-based market power” and resulting in more stringent loan conditions
such as higher borrowing costs. However, these studies do not focus on the e¤ects of bank mergers. Moreover, they often fail to control for the fact that the number of relationships may be
endogenously determined. Further, the results reported in these studies are sometimes the opposite of what theory predicts, but we cannot tell whether this is due to the possible endogeneity
of the number of relationships or whether the theoretical predictions are wrong. In contrast, by
using the event of a bank merger, the present study can measure the e¤ect of a change in the
number of relationships that is clearly exogenous. Thus, we can posit the following hypothesis
for …rms that used to transact with both of the merged banks:
Hypothesis 1: Firms that used to transact with both of the merged banks experience
an exogenous decrease in the number of relationships with banks and face more stringent loan
conditions, including higher borrowing costs, following the merger.
Second, as highlighted by Berger, Demsetz, and Strahan (1999) in their comprehensive survey, bank mergers result in a change in the organizational structures and operational procedures
of the merged banks, which a¤ects …rms through their bank relationships. Firms that had established a relationship with at least one of the merged banks are a¤ected by such changes. On the
negative side for borrower …rms, the newly merged bank may be too big and organizationally
too multi-layered to extend loans to relatively small borrowers. In addition, there may be a loss
of soft information by the bank when it consolidates its branches and reallocates loan o¢ cers
somewhere unrelated to old borrowers. The possible result is more stringent loan conditions
for borrowers. On the positive side, the newly merged bank may successfully reduce operating
costs and increase its managerial e¢ ciency. This improvement could potentially result in more
favorable transaction terms for borrowers. To summarize, we have the following hypothesis:
Hypothesis 2: Firms that used to transact with one of the merged banks also face more
stringent loan conditions due to organizational changes or the loss of soft information at the
bank. Alternatively, they face more favorable conditions if they receive rents created by improved
managerial e¢ ciency at the merged bank.
7
Third, as shown by Karceski, Ongena, and Smith (2005), bank mergers may have an asymmetric e¤ect on …rm-bank relationships depending on whether a …rm used to transact with the
acquiring bank or the acquired bank. In the case they examined, listed …rms in Norway that
transacted with acquired banks saw a decrease in their stock price following the bank merger.
Our setting here allows us to examine similarly whether the e¤ects of the bank merger are asymmetric, depending on whether a …rm used to transact with the acquired (UFJ) or the acquiring
bank (BTM).
Hypothesis 3: Firms that used to transact with the acquired bank face more stringent loan
conditions than …rms that transacted with the acquiring bank.
We are going to examine each of these three hypotheses using our …rm-level panel dataset
and employing a propensity score matching estimator. Details of the procedure are provided in
Sections 5 and 6.
3
3.1
The BTM-UFJ Merger
Japan’s banking system and merger activity
This section provides a brief description of the structure of, and developments in, Japan’s banking sector in the past few decades, including merger and acquisition activity. Roughly speaking,
banks in Japan are categorized into two groups based on the nature of their activities: major
banks that operate nation-wide and often across national borders, and regional …nancial institutions that operate in relatively limited geographical areas.1 The major banks were traditionally
further divided into city banks, trust banks, and long-term credit banks, although the latter
type no longer exists.2 The city banks are legally categorized as regular commercial banks and
are the largest in terms of the size of their assets. They extend loans not only to large …rms but
also to small businesses and individuals. The trust banks extend loans mainly to large …rms and
provide trustee services to customers. The long-term credit banks provided long-term loans to
large …rms while issuing long-term debentures in order to collect funds from the public.
1
There exist other types of banks in Japan, including agricultural/…shery cooperatives, government-a¢ liated
…nancial institutions, and de novo banks. However, we do not include these types of institutions in our brief
overview since they are of little relevance to the issues considered here.
2
There used to be three long-term credit banks; however, two of them failed around the turn of the millennium,
while the third merged with two city banks and thus became a regular commercial bank.
8
Regional …nancial institutions comprise of regional banks, second-tier regional banks, shinkin
banks, and credit cooperatives. The regional banks are regular commercial banks and the largest
in size among the regional …nancial institutions. In most cases, however, they focus on local loan
markets at the prefecture level. The second-tier regional banks, which used to be mutual banks,
primarily lend to small businesses. In 1989, they converted themselves into regular commercial
banks and started to be labelled as second-tier regional banks. They usually operate in one or
a few adjoining prefectures. Both the shinkin banks and the credit cooperatives are non-pro…t
cooperatives composed of members living and working in a de…ned geographical area. They
extend loans mainly to their members, including small businesses.
Until the 1990s, bank mergers were very rare in any of these bank categories. The number
of city banks remained unchanged at 13 until 1990. Mergers among regional and second-tier
regional banks were also rare. Only three of the second-tier banks were acquired during the
1970s and 1980s. The number of mergers involving shinkin banks and credit cooperatives was
also limited until the 1990s.
This stability in the number of banks to a considerable extent was the result of the so-called
“convoy system,”in which competition among banks was limited due to government restrictions
on the opening of new branches; competition between banks and other categories of …nancial
businesses, such as security houses and insurance …rms, was also strictly prohibited; and, against
the background of these policies, the government arm-twisted larger healthy banks into acquiring
failing banks in exchange for the permission to open new branches. As a consequence, most bank
mergers during the 1970s and 1980s were initiated at the request of the government in order to
bail out weaker banks.
However, Japan’s stable banking system became increasingly fragile in the 1990s. This was
partly due to the prolonged decline in asset prices during this period and partly due to the
…nancial liberalization undertaken in the 1980s. Yet another factor was the introduction of the
Basel Accord, which stipulated risk-based capital requirements and led relatively weak banks
to consolidate. As a consequence, there were two mergers among city banks and three mergers
among regional banks in the early 1990s.
9
3.2
A wave of bank mergers
However, starting in 1997, Japan experienced a …nancial crisis that set o¤ a veritable wave of
bank mergers. As a consequence of the crisis, triggered by non-performing loans, the Japanese
government was forced to inject large sums into the banking system, resulting in quasi-nationalization
that provided the impetus for wide-ranging consolidation in the …nancial sector. Major mergers during this phase included those between two city banks and one long-term credit bank,
Dai-ichi Kangyo Bank, Fuji Bank, and the Industrial Bank of Japan, to form Mizuho Financial
Group in 2000, and between Sakura and Sumitomo banks to form the Sumitomo Mitsui Banking
Corporation in 2001. The merger between BTM and UFJ did not follow until 2005.
3.3
The BTM-UFJ merger
The Bank of Tokyo-Mitsubishi (BTM), which acquired UFJ Bank, was itself the product of
a merger between Mitsubishi Bank, one of the largest city banks, and the Bank of Tokyo in
1996. In contrast with most other city banks, BTM had remained relatively healthy throughout
the …nancial crisis. UFJ, on the other hand, resulted from the merger in 2001 of Sanwa Bank
and Tokai Bank and had massive amounts of non-performing loans to several ailing large …rms
without su¢ cient reserves for loan losses, although this fact had not been made public at the
time. Following a severe dispute with the Financial Services Agency (FSA) on the treatment
of these loans, UFJ was forced to report huge loan losses in its …nancial statement in May
2004, meaning that its capital level was critically low. Desperately in need of cash to shore
up its balance sheet and ensure a su¢ cient level of capital, UFJ agreed with BTM to merge,
with the announcement being made in July 2004 and the merger itself becoming e¤ective as of
2005. Bringing together the second-largest (BTM) and the fourth-largest (UFJ) bank in Japan,
the merger created the largest …nancial institution not only in Japan but also in the world,
outstripping Citigroup Incorporated in terms of assets.
Several remarks concerning the merger are in order. First, the interval between the disclosure
of UFJ’s massive loan losses (May 2004) and the announcement of the merger (July 2004) was
short. Until the disclosure of those losses, most UFJ o¢ cials as well as borrower …rms and other
customers did not appear to have expected any radical changes in UFJ’s management.
10
Moreover, even after UFJ had been forced to disclose its losses, it initially did not intend to
merge with BTM. Instead, it tried to sell one of its operating arms, UFJ Trust Bank, to another
…nancial group, Sumitomo Trust Bank. Therefore, it seems fair to say that the behaviour of
neither UFJ nor borrowers was a¤ected by the expectation of a merger until the merger was
formally announced.
Second, the merger between BTM and UFJ was almost the last in the merger wave in the
Japanese banking sector. There has been no merger involving a city bank since 2005. Even in
terms of smaller mergers, there were only one second-tier bank merger and a few shinkin bank
mergers in the year 2005, while since then, only 34 regional, second-tier regional, and shinkin
banks have been involved in mergers.3 Thus, focusing on the merger between BTM and UFJ
allows us to examine the e¤ects of a bank merger without confounding factors caused by other
big mergers.
The third remark is that the loan losses that triggered the merger were due to non-performing
loans to a small number of large …rms that were considered to be “too big to fail.”The average
ex-ante performance of UFJ’s small business borrowers, on the other hand, was not signi…cantly
worse in our sample than that of BTM’s small business borrowers. Thus, it is unlikely that
UFJ’s balance sheet problems and subsequent merger with BTM were caused by the ex-ante
under-performance of UFJ’s small business borrowers.
4
Data
4.1
Data sources
The data used in this study are taken from the database of Tokyo Shoko Research (hereafter
TSR database), a business database company. The TSR database covers more than 1.2 million
…rms in Japan and provides information on …rms’ primary characteristics such as …rm age,
number of employees, ownership structure, industry, location, and the identity of banks and bank
branches the …rm transacts with. For a sizable subset of …rms, the database also has information
on …nancial statements, including the outstanding amount of assets, interest payments, the
outstanding amount of short- and long-term loans, business pro…ts, and the outstanding amount
3
Speci…cally, one regional bank, two second-tier regional banks, 19 shinkin banks, and 12 credit cooperatives
have been involved.
11
of capital. The sample we use for analysis comprises 18,888 …rms that had responded to previous
government surveys implemented annually in 2001-2003 by the Small and Medium Enterprise
Agency. Firms surveyed were randomly drawn from the TSR database. For each of the sample
…rms, we add information on their primary characteristics for 2005 and 2008. We also add
information on their balance sheets and their pro…ts and losses for the years 2004-2008.
Since our focus is on borrower …rms that are likely to be credit-constrained, we limit our
analysis to non-listed …rms in the database. Another reason for excluding listed …rms from the
analysis is that it is the large listed …rms whose underperformance resulted in the massive nonperforming loans that eventually triggered the merger of UFJ and BTM. Moreover, we restrict
the sample to …rms that transact with at least two banks in order to examine the e¤ect of an
exogenous reduction in the number of …rm-bank relationships. Further, since there were bank
mergers in 2005 other than the one between BTM and UFJ, we exclude …rms that transacted
with one or more of these other banks from the sample. As a result, we have an unbalanced
panel dataset of 11,107 …rms.
4.2
Variables
We have several sets of variables to examine our empirical hypotheses. A list of the variables
and their de…nitions is provided in Table 1.
4.2.1
Outcome variables
First, in order to gauge the availability of loans to a …rm, we use several variables. The …rst
variable is the borrowing costs a …rm faces (RAT E), calculated using information from …rms’
…nancial statements. Following Caballero, Hoshi, and Kashyap (2008), we use interest and
discount expenses divided by the sum of long-term loans, short-term loans, and bills discounted
in the previous period. We also use the log of RAT E (lnRAT E) in order to examine the
semi-elasticity of borrowing costs to changes in the explanatory variables. As for the amount
of loans that each …rm obtains, we use total loans, short-term loans, and long-term loans, all
standardized by the total amount of assets. We label these variables LOAN , SHORT , and
LON G, respectively. Finally, to represent …rms’performance, we use the return on total assets
outstanding (ROA) and the capital ratio (CAP ).
12
4.2.2
Variables used for the propensity score estimation
To examine the determinants of transaction relationships with banks, we consider four categories
of variables: …rm-bank relationship characteristics, which are used as the dependent variables,
as well as …rm characteristics, regional dummies, and industry dummies. For the …rm-bank
relationship variables, we generate several dummy variables based on the information on the
identity of the banks and bank branches that …rms transact with. In the TSR database, each
…rm reports the identities of up to ten banks and their branches that they deal with. When,
in 2005, none of these banks are BTM or UFJ, we set M ERGER0=1 and 0 otherwise. When
one of these banks is either BTM or UFJ, we set M ERGER1=1 and 0 otherwise. Moreover,
in order to determine whether the e¤ects of the merger are asymmetric depending on whether
a …rm transacted with BTM or UFJ prior to the merger, we set M ERGER1_1=1 when UFJ
is one of the banks listed by a …rm and 0 otherwise, and M ERGER1_2=1 when BTM is one
of the banks and 0 otherwise. When the banks that a …rm transacted with include both BTM
and UFJ, we set M ERGER2=1 and 0 otherwise. The variable M ERGER2 is used speci…cally
to examine our …rst hypothesis.
Summarizing these binary variables, we create two index variables, which are used for
the multinomial probit estimations. The …rst is M T Y P E1, for which the set of values is
M T Y P E1 = f0; 1; 2g. Firms with M ERGER0 = 1 have a value of zero, those with M ERGER1 =
1 have a value of one, and those with M ERGER2 = 1 have a value of two. The second index variable is M T Y P E2, which considers the acquiring and the acquired bank separately and for which
the set of values consequently is M T Y P E2 = f0; 1; 2; 3g. Firms with M ERGER0 = 1 have a
value of zero, those with M ERGER1_1 = 1 have a value of one, those with M ERGER1_2 = 1
have a value of two, and those with M ERGER2 = 1 have a value of three. The great advantage
of focusing on a merger of megabanks, as we are doing here, is that it allows us to employ this
range of dummy variables as there are a su¢ cient number of observations that fall into each
category.
As for …rm characteristics, we use …ve variables: …rm age, …rm size, credit risk, and the
number of banks the …rm transacts with. For …rm size, we use the logarithm of the number of
13
employees as well as that of the total amount of assets outstanding. For credit risk, we employ
the credit score provided in the TSR database as a proxy. The TSR credit score is an indicator
widely used by …nancial institutions and non-…nancial …rms to assess the credit risk of small
businesses in Japan. Finally, the number of banks a …rm transacts with is employed to gauge
the relative importance of a speci…c …rm-bank relationship.
In addition to the variables listed above, we also employ dummy variables for the region in
which …rms are located (10 regions) and for the industry a …rm belongs to (11 industries).
5
Empirical Approach
5.1
Propensity score matching with multiple treatments
We measure the e¤ect of the merger of BTM and UFJ on …rms that transacted with either
one or both of these banks. In most cases, we compare …rms in the treatment group that had
transaction relationships with either one (or both) of the merged banks and those in the control
group that had no transaction relationship with the merged banks.
For each of these treatment and control group …rms, we …rst calculate the di¤erences of
variables before and after the merger. Then we calculate another di¤erence, namely the di¤erence
in these di¤erences between the treatment and the control group. This estimator is the di¤erencein-di¤erences (DID) estimator. The DID estimator …rst controls for …rms’time-invariant …xed
e¤ects by taking the di¤erences of a variable. Next, it controls for macroeconomic shocks by
taking the di¤erence between these two groups. Assuming that each borrower …rm is too small
to have a¤ected the probability of the merger between BTM and UFJ, we regard the merger as
an exogenous event.
There is possibly a selection bias in the DID estimator, since the …rms in the treatment group
are often sizable and creditworthy. Further, many of the …rms in the treatment group are located
in metropolitan areas. The treatment e¤ect for …rms with such characteristics may signi…cantly
di¤er from the treatment e¤ect for …rms with di¤erent characteristics. In order to control for
the potential selection bias, we therefore employ the propensity-score-matching di¤erence-indi¤erence (PSM-DID) estimator proposed by Rosenbaum and Rubin (1983). The estimator
is unbiased for the average treatment e¤ect on the treated (ATT) under the assumptions of
14
unconfoundedness and the balancing condition.
However, for our purposes, the PSM-DID estimator as proposed by Rosenbaum and Rubin
still su¤ers from the shortcoming that it allows for only a single type of treatment, while we need
to have multiple treatment groups for our analysis: namely, a group of …rms that transacted with
both of the merged banks, a group of …rms that transacted with either one of the merged banks,
and a group of …rms that transacted with the acquired (acquiring) bank. If we put all of these in
one treatment group, we end up in confounding a variety of e¤ects and cannot tell if the e¤ects of
the merger result from the exogenous decrease in the number of …rm-bank relationships or from
the organizational change at the merged bank. In order to overcome this problem, we adopt
the PSM-DID estimator proposed by Lechner (2002), which allows for multiple treatments and
calculates propensity scores from the multinomial probit model estimation. See the Appendix
for details of how we employ the methodology proposed by Lechner.
5.2
Examination of hypotheses
Following Lechner (2002), we allow for multiple treatments and employ the multinomial probit
model in order to obtain propensity scores for each outcome. We then arbitrarily choose pairs
of outcomes fl; mg and calculate conditional propensity scores. We use the group of …rms with
outcome fmg as treatments and the group of …rms with outcome flg as controls.
In order to examine Hypotheses 1 and 2, we employ the index variable M T Y P E1 whose
values in the set f0; 1; 2g correspond to the three mutually exclusive outcomes M ERGER0 =
1; M ERGER1 = 1; andM ERGER2 = 1, implement a baseline multinomial probit estimation,
and calculate propensity scores for each outcome. We then choose three pairs of values, namely
M T Y P E1 = f0; 2g,f1; 2g, and f0; 1g. The …rst two pairs are used to examine Hypothesis 1.
Using the pair f0; 2g, we compare …rms that transacted with both BTM and UFJ and those
that transacted with neither BTM and UFJ. The di¤erence of these two outcomes, however,
includes two distinct e¤ects, namely, the e¤ect of increased switching costs and the e¤ect of
managerial changes at the merged bank. In order to isolate the former e¤ect, we employ the
pair f1; 2g, where …rms for which M ERGER2=1 are the treatment group and those for which
M ERGER1=1 are the control group.
15
The third pair is used to examine Hypothesis 2. Using the pair of outcomes f0; 1g means that
we are employing …rms that transacted with one of the merged banks as the treatment group
and …rms that transacted with neither of them as the control group. Estimating the treatment
e¤ect using our sample allows us to examine the e¤ects of the merger transmitted through the
relationship between a …rm and one of the merged banks.
In order to examine Hypothesis 3, we use the index variable M T Y P E2 whose values in
the set f0; 1; 2; 3g correspond to four outcomes that are again mutually exclusive, namely
fM ERGER0 = 1; M ERGER1_1 = 1; M ERGER1_2 = 1; M ERGER2 = 1g, and implement
a multinomial probit estimation. The di¤erence from the baseline multinomial probit estimation
is that we further divide the outcome M ERGER1 = 1 into the outcome M ERGER1_1 = 1,
in which …rms transacted with UFJ, and the outcome M ERGER1_2 = 1, in which …rms
transacted with BTM. After attaching the propensity scores based on the multinomial probit
estimation, we choose three pairs of values, namely f0; 1g, f0; 2g, and f2; 1g. Using the third pair
of outcomes, …rms that transacted with UFJ are the treatment group and …rms that transacted
with BTM are the control group. Estimating the treatment e¤ect allows us to examine if there
are any asymmetries in the way …rms are a¤ected by the merger depending on which of the two
banks they transacted with, that is, whether they transacted with the acquiring bank (BTM)
or with the acquired bank (UFJ).
6
6.1
Results
Summary statistics
In this subsection, we provide summary statistics for the variables introduced in the previous
section. Table 2 shows the summary statistics for the entire sample used for the multinomial
probit model estimation, while Table 3 shows the summary statistics for the entire sample as well
as for the subsamples that satisfy M ERGER0 = 1, M ERGER1 = 1, and M ERGER2 = 1.
Finally, Table 4 summarizes the variables used for measuring outcomes of the bank merger.
16
6.1.1
Variables used for the multinomial probit model estimation
Table 2 shows that the means of M ERGER0, M ERGER1, M ERGER2, M ERGER1_1, and
M ERGER1_2 are 0.636, 0.254, 0.110, 0.136, and 0.118, respectively, indicating that about 64%
of …rms in the entire sample did not have a transaction relationship with either of the merged
banks prior to the merger, while about 25% had a transaction relationship with one of them and
a further 11% with both. The 25% of …rms that used to transact with one of the merged banks
are relatively evenly split between those that used to transact with UFJ (14%) and those that
used to transact with BTM (12%).
The mean values of lnEM P and lnASSET S are 3.69 and 0.15, corresponding to 40.06
employees and 1.17 billion yen in real terms. The distributions of these variables are skewed to
the left, with an overwhelming majority of small …rms and a small number of large …rms that
signi…cantly increase the mean values. The mean of the proxy for …rms’credit risk, RAT IN G, is
about 56, which is above the average for all …rms in the TSR database. In terms of …rms’location,
the KAN T O area has the largest number of …rms, followed by KIN KI and T OHOKU .
Table 3 presents the summary statistics for the di¤erent subsamples. As can be seen, there
are considerable di¤erences across the subsamples in the means of many of the variables just
mentioned. The mean of F IRM AGE di¤ers moderately across subsamples: …rms that transacted with both BTM and UFJ (M ERGER2 = 1) are the oldest on average, while those that
transacted with neither of the two (M ERGER0 = 1) are the youngest. On the other hand,
there are substantial di¤erences in the variables on …rm size: …rms that transacted with both
BTM and UFJ (M ERGER2 = 1) were the largest in terms of employment with a mean of 238
employees and in terms of assets outstanding with a mean of 14 billion yen, followed by …rms that
transacted with one of the merged banks (M ERGER1 = 1), while …rms that transacted with
neither of the merged banks were the smallest with a mean of 63 employees and 3 billion yen of
total assets outstanding. Similar patterns can be found regarding the number of banks each …rm
transacted with as well as the credit score in that …rms that fell into the M ERGER2 = 1 group
have the highest values on average, followed by those falling into the M ERGER1 = 1 group.
Since BTM and UFJ were city banks and the second- and fourth-largest in Japan, respectively,
17
…rms that transacted with them were on average also larger and more creditworthy than most
other …rms.
Firms’ location also di¤ers signi…cantly across the di¤erent subsamples, which presumably
re‡ects the geographical distribution of bank branches. The bank branches of both BTM and
UFJ were concentrated in the metropolitan areas of KAN T O, KIN KI, and T OKAI, although
each bank had at least one branch in each area. Approximately half of the …rms that transacted
either with one or both of the banks (M ERGER1 = 1 or M ERGER2 = 1) are located in
KAN T O, about a quarter to a third are located in KIN KI, and between 10 and 18 percent
are located in T OKAI. These …gures are much higher than the corresponding …gures for …rms
that transacted with neither of the two (M ERGER0 = 1).
Taken together, these results suggest that there is considerable heterogeneity across the
di¤erent subsamples, which is the reason why we decided to employ the matching approach
outlined above.
6.1.2
Variables that measure the e¤ects of the bank merger
Next, we provide an overview of the variables that measure the e¤ects of the merger, including
borrowing costs, loan availability, and …rms’ex-post performance. Table 4 shows not only the
level of each outcome variable in the year of the merger (t = 2005), but also its development
from t to t + 1, t + 2, and t + 3. For borrowing costs, the mean value for the entire sample
in year t is 2.54%. Looking at the di¤erent subsamples, …rms that transacted with both banks
(M ERGER2 = 1) paid the lowest interest rates (2.39%), while those that transacted with
neither (M ERGER0 = 1) paid the highest rates (2.61%). Turning to the development in
borrowing costs from year t, for the sample as a whole, there is actually a slight decrease from t
to t + 1, followed by increases in t + 2 and t + 3. This trend re‡ects the tightening of monetary
policy, which started in March 2006. Looking at the di¤erent subsamples, …rms that transacted
with both banks (M ERGER2 = 1) experienced the largest increase in borrowing costs (+0.68
%), followed by …rms that transacted with one of the banks (M ERGER1 = 1) (+0.42%), while
…rms that transacted with neither bank (M ERGER0 = 1) saw the smallest increase (+0.17%).
Figure 1 shows that there are considerable di¤erences across the subsamples in the way the
18
distribution of borrowing costs evolves over time from year t to t + 3. Speci…cally, although the
distribution shifts to the right in all subsamples, the extent of the shift appears to be greatest for
…rms that transacted with both banks (M ERGER2 = 1) and smallest for those that transacted
with neither (M ERGER0 = 1). Even though a number of factors are yet to be controlled for,
this simple comparison between subsamples suggests that …rms that transacted with the two
merged banks tend to have experienced a larger increase in borrowing costs than …rms that had
no relationship with the two banks. We will revisit this point using the PSM-DID estimator
later in this section.
Returning to Table 4, considerable di¤erences across subsamples can also be observed for
the loan availability variables, especially for LOAN and LON G. In year t, the levels of these
variables were higher among …rms that had transacted with neither of the banks (M ERGER0 =
1) than those that had transacted with both (M ERGER2 = 1). Moreover, for …rms that had
transacted with neither of the banks (M ERGER0 = 1), LOAN three years after the merger
(t+3) was only marginally lower (-0.17%) than in the year of the merger, while LON G had even
increased (+0.76%). In contrast, …rms that transacted with one or both of the banks prior to
the merger (M ERGER1 = 1 or M ERGER2 = 1) experienced a decline in both these variables.
Finally, let us examine the …rm performance variables, return on assets (ROA) and the
capital ratio (CAP ). As for ROA, three years after the merger, …rms that had transacted with
one or both of the merged banks (M ERGER2 = 1 and M ERGER1 = 1) had experienced a
smaller decline than …rms that had transacted with neither of the banks (M ERGER0 = 1). At
the same time, the former also saw a larger increase in CAP than the latter.
6.2
Multinomial probit estimation
We proceed to estimate the multinomial probit models in order to attach propensity scores
to each observation. In our baseline model, we use M T Y P E1. The marginal e¤ects when
M T Y P E1 takes a value of 1 or 2 are shown in panel (a) of Table 5. For these values of the
dependent variable, most of the explanatory variables have signi…cant parameters. The signs
of the parameters are almost the same for these two values and are consistent with what we
observed in the summary statistics. We …nd that larger and more creditworthy …rms, as well as
19
…rms located in metropolitan areas, are more likely to have had a transaction relationship with
one or both of the banks. In addition, we …nd that the size of these parameters in most cases is
larger for M T Y P E1 = 1 than for M T Y P E1 = 2.
In addition to the baseline model, we estimated a slightly di¤erent model in which the
dependent variable is M T Y P E2. The marginal e¤ects when M T Y P E2 takes a value of 1, 2,
or 3 are shown in panel (b) of Table 5. For some of the variables we observe di¤erent parameter
values for M T Y P E2 = 1 and M T Y P E2 = 2, meaning that the characteristics of …rms that
transacted with BTM are somewhat di¤erent from those that transacted with UFJ. For example,
…rms were more likely to have transacted with BTM the greater their number of employees; on
the other hand, for UFJ, no such link between employment size and the likelihood of having
a transaction relationship can be observed. In addition, the signs of the parameters on several
of the area dummies di¤er for …rms that transacted with BTM and those that transacted with
UFJ. Being located in T OKAI has a signi…cant positive marginal e¤ect on the likelihood that a
…rm will have transacted with UFJ, but a signi…cant negative marginal e¤ect on the likelihood
that it will have transacted with BTM. Moreover, being located in KAN T O has a higher
marginal e¤ect on the likelihood that a …rm will have transacted with BTM than that it will
have transacted with UFJ, while the opposite is the case for KIN KI. These di¤erences in the
regional parameters between the banks may re‡ect di¤erences in the geographical distribution
of bank branches.
Using the results of the above two multinomial probit model estimations, we form several
pairs of outcomes in order to attach conditional propensity scores following Lechner’s methodology. We then estimate the treatment e¤ects. We detail these procedures in the next subsection.
6.3
Treatment e¤ect estimation
In this subsection, we estimate the treatment e¤ects of the bank merger using PSM-DID estimation. Since we allow for multiple treatments, we choose a pair of outcomes from the multinomial
probit estimation in order to match treatment observations with non-treatment observations.
20
6.3.1
Examining Hypotheses 1 and 2
We employ the following three pairs of outcome values, M T Y P E1 = f0; 2g, f1; 2g, and f0; 1g,
and estimate the treatment e¤ect within the group of …rms that take either one of the values in
each pair. Taking the …rst pair of outcomes, M T Y P E1 = f0; 2g, as an example, …rms whose
outcome value corresponds to the second value in the bracket (2 in this case, indicating that a
…rm transacted with both banks) form the treatment group, while …rms whose outcome value
corresponds to the …rst value (0 in this case, indicating that a …rm transacted with neither
of the merged banks) form the non-treatment group. We calculate the conditional propensity
scores for …rms that belong to the treatment and non-treatment groups. We then apply the
caliper matching procedure in order to choose a control group observation for each treatment
observation. We conduct PSM-DID estimation for …rms that belong to the treatment and
control groups and compare the development in outcome variables over time between these
groups. Table 6 shows the results.
Let us start with column (1), which focuses on …rms that transacted with both banks and
experienced an exogenous decrease in the number of …rm-bank relationships as a result of the
merger. In the column, we observe signi…cant increases in borrowing costs (RAT ES) for treatment …rms relative to the control …rms between years t and t + 3. This provides support for
Hypothesis 1, which stated that …rms that transacted with both BTM and UFJ were expected
to face higher switching costs and more stringent borrowing conditions, including higher interest
rates, following the merger. Column (1) shows that RAT E increased by 40 basis points (bp)
for …rms that had transacted with both banks relative to …rms that had transacted with neither
bank.
Next, turning to Column (3), it is noteworthy that even here we …nd signi…cant treatment
e¤ects for RAT E. Since …rms in the treatment and the control group have in common that
they had a transaction relationship with the newly merged bank and they only di¤er in that the
former transacted with both banks while the latter transacted with only one, one might expect
that they were a¤ected by the merger in the same way. Yet, this is not the case. What this
suggests is that the merger reduced the number of …nancing sources for treatment …rms (…rms
21
that transacted with both banks prior to the merger), resulting in a more stringent environment
for the procurement of funds for these …rms, while the number of …nancing sources for control
…rms remained unchanged.4
Further, Column (5) of Table 6 examines Hypothesis 2 that the procurement conditions
of …rms that transacted with at least one of the merged banks were adversely a¤ected due to
organizational changes at the merged bank. Possible adverse e¤ects include the loss of soft
information through the consolidation of branch networks and the reallocation of loan o¢ cers
who have established long-term relationships with borrowers. On the other hand, it is also
possible that …rms that transacted with at least one of the merged banks before the merger may
in fact enjoy more favorable procurement conditions after the merger, for example because of
e¢ ciency improvements at the merged bank. The results in the table, however, suggest that the
negative e¤ects dominate: relative to control …rms, treatment …rms’borrowing costs (RAT E)
three years after the merger had increased by 20bp.
A few additional remarks regarding the treatment e¤ects obtained in the PSM-DID estimation in Table 6 are in order. First, essentially no statistically signi…cant treatment e¤ects can
be observed for any of the variables apart from RAT E and lnRAT E. Although based on the
treatment e¤ects calculated using simple DID estimation shown in Table 4 one might expect
several other variables, such as LOAN and LON G, to be signi…cant, this is not the case in the
PSM-DID estimation. Moreover, the maturity composition of loans in terms of SHORT and
LON G loans as well as …rm quality as measured by ROA and CAP show no signi…cant change
after the merger. Therefore, we are able to say that the increase in RAT E and lnRAT E seen
in Table 6 is not due to a shift from short-term to long-term loans or to a change in the quality
of borrower …rms.
Second, the results in Table 6 suggest that it takes about three years for RAT E and lnRAT E
to increase by a statistically signi…cant margin. However, this does not necessarily mean that
most individual loan contracts were …xed for the …rst year or two after the merger. Our measure
of borrowing costs is the amount of annual interest payments divided by the total amount of
4
Whether this result is a re‡ection of the Japanese context, where …rm-bank relationships tend to be stable
and establishing a new relationship is relatively costly, or whether a similar pattern can be observed elsewhere, is
an interesting topic for future research.
22
loans outstanding. It may well be the case that individual loan contracts were revised to higher
interest rates immediately after the bank merger. However, it may take some time for these new
contracts to make up a substantial share in the total amount of loans outstanding.
6.3.2
Examining Hypothesis 3
We now turn to the examination of Hypothesis 3, which states that …rms that transacted with the
acquired bank are expected to face more stringent loan conditions than …rms that transacted with
the acquiring bank. We form the following three pairs of outcome values: M T Y P E2 = f0; 1g,
f0; 2g, and f2; 1g. Firms whose outcome value corresponds to the second value in the bracket
belong to the treatment group and …rms whose outcome value corresponds to the …rst value
belong to the non-treatment group. Following the same steps as in the baseline case, we calculate
conditional propensity scores, apply the caliper matching procedure to obtain a set of treatment
and control observations, and obtain PSM-DID estimates for the treatment e¤ects. Table 7
shows the results.
Looking at the results for RAT E three years after the merger (t+3) suggests that the positive
treatment e¤ects are very similar irrespective of whether a …rm transacted with the acquiring or
the acquired bank. Speci…cally, relative to …rms that transacted with neither of the two banks,
…rms that transacted with UFJ saw an increase in RAT E of 19bp, while the corresponding
…gure for …rms that transacted with BTM is 22bp. Thus, it seems that the newly merged bank
did not treat former UFJ borrowers in a discriminatory manner in terms of borrowing costs.
This contrasts with the results reported by Karceski, Ongena, and Smith (2005), who, focusing
on listed …rms, found that the share prices of …rms that had borrowed from the acquired bank
underperformed signi…cantly relative to those of …rms that had borrowed from the acquiring
bank.
7
7.1
Further Examinations
E¤ects of the number of banks
In addition to the examination of Hypotheses 1 through 3, we provide further empirical analyses.
In the …rst two subsections we implement robustness checks and in the third we interpret the
23
results in the context of the literature arguing that loan pricing by Japanese banks during the
1990s and early 2000s was distorted (e.g., Peek and Rosengren 2005, Caballero, Hoshi, and
Kashyap 2008). To this end, we employ a parametric approach for examining the treatment
e¤ect of the bank merger in place of the non-parametric approach with the PSM estimator used
in the previous section.
Both the parametric and non-parametric estimators for the treatment e¤ects are consistent
under the assumption of unconfoundedness and the overlap condition. Since we introduce interaction terms and allow for the treatment e¤ect to vary across subsamples, we employ the
parametric estimator - Flores and Mitnik (2009) call it the partial mean linear estimator rather than the non-parametric estimator, which is more ‡exible but relatively computationally
burdensome.
We show that this parametric partial mean linear estimator provides quantitatively similar
results to those in the baseline case presented in Table 6. We focus on the development of interest
rates, that is, RAT E and lnRAT E, between t and t + 3, since we found in Section 6 that these
variables changed signi…cantly after the merger. The equation for the baseline estimation is:
E [ RAT Eit+3 jXi ; M T Y P E1i ] = const +
X
k 1(M T Y
0
P E1i = k) + Xi
(1)
k2f1;2g
where Xi is a vector of the explanatory variables employed in the previous section and 1( ) is
an indicator variable that is unity if the condition in parentheses is satis…ed and zero otherwise.
Our focus is on the parameters
2
1,
and
1
k
with k 2 f1; 2g, which represent the treatment e¤ect.
2,
respectively represent the treatment e¤ects using the following pairs of outcome
values: M T Y P E1 = f0; 2g, f1; 2g, and f0; 1g.
The results are shown in Table 8(a), with the column numbers corresponding to those in
Table 6. The results are quantitatively similar to those for the non-parametric estimation in
Table 6: The estimator values for RAT E are 41, 21, and 20bp, while the corresponding …gures
in Table 6 are 40, 25, and 20bp, respectively, with similar levels of statistical signi…cance. The
estimates for lnRAT E are also quantitatively close to each other in Tables 6 and 8(a). We start
from this baseline parametric speci…cation and add interaction terms to examine three issues.
24
The …rst issue concerns the measurement of the interest rate paid by …rms. While Sapienza
(2002) and Erel (2011) employ contract-level interest rates, our interest rate variable is calculated
as a …rm’s total annual interest payments divided by its total amount of loans outstanding. This
means that our interest rate measure includes not only interest payments to BTM and/or UFJ,
but also to other banks, and it is therefore di¢ cult to isolate the e¤ects of the pricing behaviour
of the two banks.
One way to circumvent this problem of confounding factors with regard to the cost of borrowing is to control for the number of banks a …rm transacts with at the time of the merger. If a
…rm in the treatment group for which M T Y P E1 = 2 (i.e., it had transaction relationships with
both BTM and UFJ) had only two banks as transaction partners, then we know that after the
merger it paid interest only to the newly merged bank. Further, if the number of banks a …rm
transacts with is greater than two but nevertheless small, the merged bank will still account for
a large share of the …rm’s interest payments and the …rm is therefore more likely to be a¤ected
by the merger than other treatment …rms that transact with a large number of banks. Based
on this line of reasoning, we set up the following equation for the expected value of the change
in RAT E conditional on the existence of a relationship with one or both of the merged banks,
the number of banks, and the interaction term between the two:
E [ RAT Eit+3 jXi ; M T Y P E1i ; N BAN Ki ] = const +
+
X
k 1(
)N BAN Ki +
k2f1;2g
X
k 1(
)N BAN Ki2
X
k 1(M T Y
P E1i = k)
k2f1;2g
0
+ '1 N BAN Ki + '2 N BAN Ki2 + X(2)
i
k2f1;2g
Note that we include quadratic terms for N BAN K in order to incorporate any possible
non-linear relationships between RAT E and N BAN K. Our aim is to measure the treatment
e¤ects represented by
k,
k,
and
k.
When we compare …rms that transacted with both of the
merged banks and those that transacted with neither of them, the treatment e¤ect is calculated
as
2
+
2 N BAN K
+
2 N BAN K
2.
When we compare …rms that transacted with either one
of the merged banks and those that transacted with neither of them, the treatment e¤ect is
calculated as
1
+
1 N BAN K
+
1 N BAN K
2.
The results for the treatment e¤ect, which di¤ers depending on the number of banks each
…rm transacts with, are presented in Figure 2. In both cases, the size of the treatment e¤ect
25
gradually decreases as N BAN K increases.5 This negative correlation between the size of the
treatment e¤ect and the number of banks is consistent with our discussion above in that the
merger a¤ects loan conditions more severely for …rms with fewer alternative …nancing sources
other than the merged banks.
7.2
E¤ects of bank branch consolidation
The second issue to be examined is the role of branch consolidation in relation to the results for
Hypothesis 2. In the previous section, we speculated that the increase in borrowing costs among
…rms that transacted with the merged banks is attributable to the consolidation of the branch
network of the merged banks and the resulting loss of soft information. In order to examine
whether this is correct, we focus on …rms that transacted with branches that were closed or
merged with other branches within a few years of the merger. Following the merger between
BTM and UFJ, about 20% of all branches were closed or merged with other branches and thus
there is a su¢ cient number of …rms that had transaction relationships with such branches.
When we estimate the treatment e¤ect of bank branch consolidation, …rms that belong to
the treatment group are those that had transacted either with one or both of the merged banks
(i.e., …rms for which M T Y P E1 = 1 or 2) and that saw at least one of the branches they used to
transact with closed after the merger (CN SL = 1). Thus, we estimate the following equation
for the expected value of the change in RAT E conditional on the existence of a relationship
with one or both of the merged banks and the indicator of bank branch consolidation:
E [ RAT Eit+3 jXi ; M T Y P E1i ; CN SLi ] = const +
X
k 1(M T Y
P E1i = k)
k2f1;2g
0
+ [1(CN SLi = 1jM T Y P E1i = 1) + 1(CN SLi = 1jM T Y P E1i = 2)] + Xi
Our aim is to measure the treatment e¤ects represented by
k
and
(3)
. When we compare
…rms that transacted with at least one of the merged banks and those that transacted with
neither of them, the treatment e¤ect is calculated as
k
+ . Table 8(b) shows the estimation
5
Note that due to the small number of observations for …rms transacting with only two banks, the treatment
e¤ect for N BAN K = 2 is not signi…cantly di¤erent from zero. Note also that …rms for which N BAN K = 1
are excluded from the sample since p(M T Y P E1 = 2jN BAN K = 1) = 0 and thus these …rms do not satisfy the
positive support condition for M T Y P E1 = 2.
26
results for the treatment e¤ect of …rms’bank branch being closed, which we compare with the
results in Table 8(a). The comparison shows that the treatment e¤ect between t and t + 3 with
regard to RAT E is larger for …rms that saw their bank branch closed than for treatment …rms
in the benchmark case. For example, among treatment …rms that transacted with both of the
merged banks (i.e., M T Y P E1 = 2), the RAT E increase is 41bp in the baseline case. On the
other hand, in the case where such …rms saw their bank branch closed, the RAT E increase
becomes even larger, ranging from 56 to 68bp. We obtain similar results for the treatment e¤ect
among …rms that transacted with one of the two banks (i.e., for which M T Y P E1 = 1).
In sum, …rms that saw a bank branches they used to transact with being closed paid higher
interest rates than …rms that were only indirectly a¤ected by the consolidation of bank branches.
This indicates that the observed interest rate increase is at least partly attributable to the loss
of soft information caused by branch consolidation.
7.3
Loan pricing and credit risk
Finally, the third issue we examine is whether loan pricing following the merger better re‡ects
borrowers’credit risk. Studies on loan mispricing in Japan, such as Smith (2003) and Caballero,
Hoshi, and Kashyap (2008), have shown that during the 1990s and early 2000s the interest rates
charged by banks were insensitive to any increase in borrowers’credit risk. Our dataset allows
us to examine whether more recently the interest rates charged by banks have come to better
re‡ect borrowers’credit risk, so that riskier …rms are charged signi…cantly higher interest rates
than before, while less risky …rms are not.
Thus, we estimate the following equation for the expected value of the change in RAT E
conditional on the existence of a relationship with one or both of the merged banks and an
indicator of riskiness, RISKY . We de…ne RISKY in two ways. In the …rst de…nition, it is set
to unity if a …rm’s RAT IN G is below the median value, and zero otherwise. In the second, it
is set to unity if the …rm’s interest rate spread is negative, and zero otherwise.6 We examine if
treatment …rms for which RISKY = 1 experience a greater increase in borrowing costs after the
bank merger than …rms for which RISKY = 0. Speci…cally, we estimate the following equation:
6
This is the de…nition employed by Caballero, Hoshi, and Kashyap (2008) to detect “zombie” …rms.
27
E [ RAT Eit+3 jXi ; M T Y P E1i ; RISKYi ] = const +
+
X
k 1(M T Y
X
k 1(M T Y
P E1i = k)
k2f1;2g
0
P E1i = k)1(RISKYi = 1) + Xi
(4)
k2f1;2g
Our aim is to measure the treatment e¤ects represented by
k
and
k,
that is,
k + k.
Table
8(c) shows the estimation results for the treatment e¤ects. Relative to the treatment e¤ects in
the baseline case (Table 8(a)), however, Table 8(c) indicates that riskier …rms did not necessarily
face a larger increase in borrowing costs. Thus, although borrowing costs rose signi…cantly after
the bank merger, we …nd no evidence that loan pricing after the bank merger better re‡ects
borrowers’credit risk.
8
Conclusion
This study examined the e¤ects of a major bank merger on …rms’ …nancing conditions by focusing on Japan’s largest bank merger in history, that between the Bank of Tokyo-Mitsubishi
(BTM) and UFJ Bank (UFJ) in 2005. In contrast with many previous studies investigating the
e¤ects of bank mergers, including those by Sapienza (2002) and Erel (2011), which concentrate
on the impact on local loan markets, the present study focused on the role of …rm-bank relationships in transmitting the e¤ects of a bank merger. This emphasis on …rm-bank relationships is
based on the theoretical literature on …nancial intermediation (e.g., Sharpe 1990; Rajan 1992),
which assumes that banks establish customer relationships with borrowers in order to gather
information and that such information is available only to banks that have lent to a …rm. In
this case, it is unlikely that both incumbent and potential lenders are able to exercise market
power and set loan conditions such as interest rates in a similar manner. Therefore, …rm-bank
relationships play an important role in determining loan conditions.
The megabank merger we used for our analysis provides an excellent case study for examining
the relevance of such …rm-bank relationships, since both of the merged banks had relationships
with a large number of …rms. In addition, a substantial number of …rms had relationships with
both banks at the time of the merger and continued to maintain them for a considerable period
of time. Exploiting the information on …rm-bank relationships in our dataset, we were able to
28
investigate how the impact of the merger on …rms’borrowing conditions di¤ered depending on
whether they had a transaction relationship with none, one, or both of the merged banks. Our
…ndings can be summarized as follows.
First, the borrowing costs of …rms that transacted with both BTM and UFJ prior to the
merger increased by 40bp relative to …rms that transacted with neither. We also detect significant treatment e¤ects even after controlling for the number of banks a …rm transacted with.
Second, the borrowing costs of …rms that transacted with one of the two banks rose by a smaller
but still signi…cant margin of 20bp. The margin increases to 28-36bp when we focus on …rms that
transacted with bank branches that were subsequently closed. Third, we do not …nd a signi…cant
di¤erence in the extent of the relative increase in borrowing costs between …rms that transacted
with the acquiring bank (BTM) and those that transacted with the acquired bank (UFJ). Overall, these results suggest that the bank merger increased …rms’borrowing costs partly through
an exogenous decrease in the number of …rm-bank relationships and partly through changes in
the merged bank’s organizational structure. Finally, the treatment e¤ect does not signi…cantly
di¤er across …rms of di¤erent degrees of riskiness, meaning that our results provide no evidence
that the interest rate that a …rm is charged better re‡ects …rms’credit risk than before.
There are several directions for future research. First, the newly merged bank may have
“dropped”particular …rms. Our analysis here is limited to …rms that maintained their relationship with the newly merged bank during the period 2005-2008. However, examining what …rms
were dropped and why following the bank merger and comparing the results with those of other
studies such as the ones by Degryse, Masschelein, and Mitchell (2011) and Di Patti and Gobbi
(2007) represents an interesting avenue for further research.
Second, it would be worthwhile to analyze the impact of bank mergers on the Japanese loan
market from a more comprehensive perspective rather than just focusing on one speci…c merger.
Whether similar patterns can be found as in this study, i.e., that borrowers tend to face higher
borrowing costs following a merger, has important policy implications.
29
A
Appendix: Propensity Score Matching Estimation with Multiple Treatments
The treatment e¤ect of the merger we would like to detect is the average treatment e¤ect on
the treated (ATT), which is expressed as
AT T
= E (Y (1) j M T Y P E = 1)
E (Y (0) j M T Y P E = 1)
(5)
A simple comparison of the outcome variables for …rms that transacted with a bank that merged
(M T Y P E = 1) and those for …rms that did not (M T Y P E = 0) can be biased. More precisely,
if outcomes of Y (0) are expected to be di¤erent between …rms that transacted with the merged
bank and those that transacted with neither of the merged banks, the simple comparison has
the following bias:
E (Y (1) j M T Y P E = 1)
=
AT T
E (Y (0) j M T Y P E = 0)
+ E (Y (0) j M T Y P E = 1)
E (Y (0) j M T Y P E = 0)
(6)
To circumvent this problem, we need to control for possible selection bias in our estimation.
Thus, we employ the propensity score matching (PSM) estimation approach proposed by Rosenbaum and Rubin (1983). Their methodology is applicable to the case in which the treatment
is a binary choice. However, in practice, choices often are multinomial rather than binary. For
example, among the …rms that transacted with the merged banks, there is likely to be heterogeneity regarding the way they were involved with the banks that merged. That is, some …rms
will have transacted with the acquiring bank only, while others will have transacted with the
acquired bank only, and yet others will have transacted with both banks. Since each of these
treatment groups potentially faces di¤erent outcomes from the bank merger, it is necessary to
examine the di¤erences among the di¤erent treatment groups.
Lechner (2002) extends the analysis of Rosenbaum and Rubin, allowing for multiple treatments. In our case, we de…ne the set of treatments as M T Y P E = f0; 1; : : : ; M g, where M
2.
The corresponding outcomes for these treatments are fY (0); Y (1); : : : ; Y (M )g. Unconfoundedness is assumed as
30
fY (0); Y (1); :::; Y (M )g ? M T Y P E j X
(7)
Further, another assumption, which we call the balancing condition, has to be satis…ed in order
to ensure that we have a consistent estimator of the treatment e¤ect,
X ? M T Y P Ejp(X)
(8)
In other words, for a given propensity score, there exists a pool of treatment and control observations. They are, on average, identical and the treatment observations are randomly chosen
from the pool. As Flores and Mitnik (2009) point out, satisfying the balancing condition is
more di¢ cult in the case of multiple treatments than in the case of a single treatment. Hence,
it is important to check for the existence of overlaps prior to the matching estimation. To do
so, we not only examine the distributions of propensity scores (results not shown in the paper)
but also employ the caliper matching rule, which is the most suitable approach for this purpose.
Caliper matching arbitrarily sets a tolerance level and for each treatment observation i searches
a control observation j that satis…es the condition c(pi ) = minj kpi
pj k
". For our analysis
we do not use treatment observations for which we cannot …nd a matched observation satisfying
the above condition. We set " = 0:01 here. Thus, we are more likely to satisfy the balancing
condition by employing caliper matching.
We estimate the multinomial probit model for the probability of each treatment pk
M
.
k=0
Then we calculate the probability for the treatment m conditional on a pair of two treatments
fl; mg:
pmjl;m (X) = p(M T Y P E = m j M T Y P E = l; orM T Y P E = m; X)
=
pm (X)
pl (X) + pm (X)
(9)
We employ the propensity score matching di¤erence-in-di¤erences (PSM-DID) approach.
Under the above assumptions, ATT is expressed as:
31
l;m
AT T
= Epmjl;m (X)jM T Y P E=m
E
Y (m) j pmjl;m (X); M T Y P E = m
E
Y (l) j pmjl;m (X); M T Y P E = l
(10)
= E ( Y (m) j M T Y P E = m)
h n
oi
Epmjl;m (X)jM T Y P E=m E
Y (l) j pmjl;m (X); M T Y P E = l
And a consistent PSM-DID estimator for ATT is
2
X
1
l;m
^
4 Yit+k (m)
AT T =
NT
i2fM T Y P E=mg
X
i2fM T Y P E=lg
3
w(i; j) Yjt+k (l)5
(11)
(12)
Using this estimator, we take into consideration the heterogeneity in the way …rms were involved
with the banks that merged and examine how this heterogeneity a¤ects ex-post …rm-bank relationships as well as …rms’borrowing conditions after the merger.
32
References
Angelini, P., R. Di Salvo, and G. Ferri (1998). “Availability and Cost of Credit for Small Businesses: Customer Relationships and Credit Cooperative,” Journal of Banking and Finance,
22, 925-954.
Berger, A.N., R.S. Demsetz, and P.E. Strahan (1999). “The Consolidation of the Financial Services Industry: Causes, Consequences, and Implications for the Future,” Journal of Banking
and Finance, 23, 135-194.
Berger, A.N., and T.H. Hannan (1989). “The Price-Concentration Relationship in Banking,”
Review of Economics and Statistics, 71, 291-299.
Caballero, R.J., T. Hoshi, and A. K. Kashyap (2008). “Zombie Lending and Depressed Restructuring in Japan,” American Economic Review, 98, 1943-77.
Degryse, H., M. Kim, and S. Ongena (2008). Microeconometrics of Banking: Methods, Applications, and Results, Oxford University Press, Oxford and New York.
Degryse, H., N. Masschelein and J. Mitchell (2011). “Staying, Dropping, or Switching: The
Impacts of Bank Mergers on Small Firms,” Review of Financial Studies, 24, 1102-1140.
Degryse, H., and P. Van Cayseele (2000). “Relationship Lending within a Bank-Based System:
Evidence from European Small Business Data,” Journal of Financial Intermediation, 9, 90109.
Diamond, D.W. (1984). “Financial Intermediation and Delegated Monitoring,” Review of Economic Studies, 51, 393-414.
Di Patti, B., and E.G. Gobbi (2007). “Winners or Losers? The E¤ects of Banking Consolidation
on Corporate Borrowers,” Journal of Finance, 62, 669-695.
Erel, I. (2011). “The E¤ect of Bank Mergers on Loan Prices: Evidence from the United States,”
Review of Financial Studies, 24, 1068-1101.
33
Flores, C.A., and O.A. Mitnik (2009). “Evaluating Nonexperimental Estimators for Multiple
Treatments: Evidence from Experimental Data,” IZA Discussion Paper, No. 4451.
Harho¤, D., and T. Körting (1998). “Lending Relationships in Germany: Empirical Evidence
from Survey Data,” Journal of Banking and Finance, 22, 1317-1353.
Horiuchi, A., F. Packer, and S. Fukuda (1988). “What Role Has the ’Main Bank’ Played in
Japan,” Journal of the Japanese and International Economies, 2, 159-180.
Jones, K.D., and T. Critch…eld (2005). “Consolidation in the U.S. Banking Industry: Is the
‘Long, Strange Trip’About to End?,” FDIC Banking Review, 17, 31-61.
Karceski, J., S. Ongena, and D.C. Smith (2005). “The Impact of Bank Consolidation on Commercial Borrower Welfare,” Journal of Finance, 60, 2043-2082.
Lechner, M. (2002). “Program Heterogeneity and Propensity Score Matching: An Application
to the Evaluation of Active Labor Market Policies,” Review of Economics and Statistics, 84,
205-220.
Ongena, S., and D.C. Smith. “The Duration of Bank Relationships,”Journal of Financial Economics, 61, 449-475.
Prager, R.A., and T.H. Hannan (1999). “Do Substantial Horizontal Mergers Generate Signi…cant
Price E¤ects? Evidence from the Banking Industry,” Journal of Industrial Economics, 46,
433-452.
Peek, J., and E.S. Rosengren (2005). “Unnatural Selection: Perverse Incentives and the Misallocation of Credit in Japan,” American Economic Review, 95, 1144-1166.
Petersen, M.A., and R.G. Rajan (1994). “The Bene…ts of Lending Relationships: Evidence from
Small Business Data,” Journal of Finance, 49, 3-37.
Rajan, R. G., (1992). “Insiders and Outsiders: The Choice between Informed and Arm’s-Length
Debt,” Journal of Finance, 47, 1367-1400.
34
Rosenbaum, P., and D. Rubin (1983). “The Central Role of the Propensity Score in Observational
Studies for Causal E¤ects,” Biometrika, 70, 41-55.
Sakai, K., I. Uesugi, and T. Watanabe (2010). “Firm Age and the Evolution of Borrowing Costs:
Evidence from Japanese Small Firms,” Journal of Banking and Finance, 34, 1970-1981.
Sapienza, P. (2002). “The E¤ect of Banking Mergers on Loan Contracts,” Journal of Finance,
57, 329-367.
Sharpe, S.A. (1990). “Asymmetric Information, Bank Lending, and Implicit Contracts: A Stylized Model of Customer Relationships,” Journal of Finance, 45, 1069-87.
Smith, D.C. (2003). “Loans to Japanese Borrowers,”Journal of the Japanese and International
Economies, 17, 283–304.
Uchida, H., G.F. Udell, and W. Watanabe (2008). “Bank Size and Lending Relationships in
Japan,” Journal of the Japanese and International Economies, 22, 242-267.
Weinstein, D.E., and Y. Yafeh (1998). “On the Costs of a Bank-Centered Financial System:
Evidence from the Changing Main Bank Relations in Japan,” Journal of Finance, 53, 635672.
Williamson, O. (1968). “Economies as an Antitrust Defense: The Welfare Trade-o¤,”American
Economic Review, 58, 18-36.
35
Table 1: List of variables and their definitions
Variable
Definition
Variables used for propensity score estimation
Dependent variable
MERGER0
1 if the firm had a relationship with neither BTM nor UFJ in 2005, otherwise 0.
MERGER1
1 if the firm had a relationship with either BTM or UFJ in 2005, otherwise 0.
MERGER1_1
1 if the firm had a relationship with UFJ in 2005, otherwise 0.
MERGER1_2
1 if the firm had a relationship with BTM in 2005, otherwise 0.
MERGER2
1 if the firm had a relationship with both BTM and UFJ in 2005, otherwise 0.
MTYPE1
0 if MERGER0=1, 1 if MERGER1=1, 2 if MERGER2=1
MTYPE2
0 if MERGER0=1, 1 if MERGER1_1=1, 2 if MERGER1_2=1, 3 if MERGER2=1
Firms' characteristics
l nFIRMAGE
Natural log of firm age in 2005.
l nEMP(t‐1)
Natural log of number of employees in 2004.
l nASSETS(t‐1)
NBANK
Natural log of total assets in 2004.
RATING
Credit rating in 2005 taking a value from 0 to 100.
Number of bank relationships in 2005.
Firms' location
HOKKAIDO
1 if the firm is located in the Hokkaido area, otherwise 0.
TOHOKU
1 if the firm is located in the Tohoku area, otherwise 0.
KANTO
1 if the firm is located in the Kanto area, otherwise 0.
KOSHINETSU
1 if the firm is located in the Koshinetsu area (Niigata, Nagano, and Yamanashi), otherwise 0.
HOKURIKU
1 if the firm is located in the Hokuriku area (Ishikawa, Toyama, and Fukui), otherwise 0.
TOKAI
1 if the firm is located in the Tokai area (Aichi, Shizuoka, and Gifu), otherwise 0.
KINKI
1 if the firm is located in the Kinki area, otherwise 0.
CHUGOKU
1 if the firm is located in the Chugoku area, otherwise 0.
SHIKOKU
1 if the firm is located in the Shikoku area, otherwise 0.
KYUSHU
1 if the firm is located in the Kyushu area, otherwise 0.
Industry dummies
INDUSTRY1‐11
1: Mining, 2: Construction, 3: Manufacturing, 4: Electricity, gas, and heat supply, 5:
Telecommunications, 6: Transportation, 7: Wholesale trade, 8: Retail trade, 9: Finance and
insurance, 10: Restaurants and accommodation, 11: Other
Outcome variables
RATE
Borrowing cost: interest and discount expenses divided by the sum of long‐term loans, short‐
term loans, and "notes discounted.
l nRATE
Natural log of RATE.
LOAN
Loan ratio: sum of long‐term loans and short‐term loans divided by total assets.
SHORT
Short‐tem loan ratio: short‐term loans divided by total assets.
LONG
Long‐tem loan ratio: short‐term loans divided by total assets.
ROA
Return on assets: operating profit divided by total assets.
CAP
Capital ratio: net assets divided by total assets.
36
Table 2: Summary statistics for the entire sample
Variable
Dependent variable
MERGER0
MERGER1
MERGER1_1
MERGER1_2
MERGER2
Firms' characteristics
lnFIRMAGE
lnEMP(t‐1)
lnASSETS(t‐1)
NBANK
RATING
Firms' location
HOKKAIDO
TOHOKU
KANTO
KOSHINETSU
HOKURIKU
TOKAI
KINKI
CHUGOKU
SHIKOKU
KYUSHU
Industry dummies
INDUSTRY1
INDUSTRY2
INDUSTRY3
INDUSTRY4
INDUSTRY5
INDUSTRY6
INDUSTRY7
INDUSTRY8
INDUSTRY9
INDUSTRY10
INDUSTRY11
Obs.
Mean
Median
Std. Dev.
11107
11107
11107
11107
11107
0.6358
0.2543
0.1360
0.1183
0.1099
1.0000
0.0000
0.0000
0.0000
0.0000
0.4812
0.4355
0.3428
0.3230
0.3128
11107
11107
11107
11107
11107
3.4915
3.6903
0.1545
4.2788
55.7009
3.6376
3.5835
0.0846
4.0000
55.0000
0.5717
1.2813
1.6296
2.0028
7.2863
11107
11107
11107
11107
11107
11107
11107
11107
11107
11107
0.0711
0.0972
0.2615
0.0773
0.0431
0.0952
0.1499
0.0798
0.0421
0.0827
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.2570
0.2963
0.4394
0.2671
0.2031
0.2935
0.3570
0.2709
0.2009
0.2755
11107
11107
11107
11107
11107
11107
11107
11107
11107
11107
11107
0.0071
0.2173
0.2744
0.0063
0.0275
0.0289
0.1761
0.0978
0.0435
0.0086
0.1125
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0840
0.4124
0.4462
0.0791
0.1634
0.1675
0.3809
0.2970
0.2040
0.0926
0.3160
37
Min.
1.0986
0.0000
‐5.4230
2.0000
9.0000
Max.
4.4773
8.3885
6.4185
10.0000
84.0000
Table 3: Summary statistics for subsamples
Firms' characteristics
FIRMAGE
EMP(t‐1)
ASSETS(t‐1)
NBANK
RATING
Firms' location
HOKKAIDO
TOHOKU
KANTO
KOSHINETSU
HOKURIKU
TOKAI
KINKI
CHUGOKU
SHIKOKU
KYUSHU
No. of Obs.
ALL
MERGER0
Subsample
MERGER1
MERGER2
37.4727
100.3525
5.2278
4.2788
55.7009
35.2635
62.8859
2.5951
3.7186
54.8183
40.3647
134.6686
8.0644
4.8757
56.7192
43.5610
237.6831
13.8948
6.1384
58.4505
0.0711
0.0972
0.2615
0.0773
0.0431
0.0952
0.1499
0.0798
0.0421
0.0827
0.1028
0.1459
0.1553
0.1155
0.0595
0.0586
0.0872
0.1032
0.0586
0.1133
0.0191
0.0156
0.4150
0.0149
0.0170
0.1831
0.2337
0.0478
0.0163
0.0375
0.0082
0.0049
0.5201
0.0008
0.0090
0.1032
0.3186
0.0180
0.0066
0.0106
11107
7062
38
2824
1221
Table 4: Summary statistics for the level and development of outcome variables ALL
RATE(t)
ΔRATE(t+1)
ΔRATE(t+2)
ΔRATE(t+3)
CASH(t)
ΔCASH(t+1)
ΔCASH(t+2)
ΔCASH(t+3)
LOAN(t)
ΔLOAN(t+1)
ΔLOAN(t+2)
ΔLOAN(t+3)
SHORT(t)
ΔSHORT(t+1)
ΔSHORT(t+2)
ΔSHORT(t+3)
LONG(t)
ΔLONG(t+1)
ΔLONG(t+2)
ΔLONG(t+3)
ROA(t)
ΔROA(t+1)
ΔROA(t+2)
ΔROA(t+3)
CAP(t)
ΔCAP(t+1)
ΔCAP(t+2)
ΔCAP(t+3)
0.0254
‐0.0004
0.0012
0.0030
0.1761
‐0.0027
‐0.0048
‐0.0083
0.3532
‐0.0057
‐0.0107
‐0.0083
0.1535
0.0125
0.0095
‐0.0017
0.2290
0.0002
0.0000
0.0006
0.0240
0.0002
‐0.0014
‐0.0034
0.2814
0.0039
0.0064
0.0123
Subsample
MERGER0=1
MERGER1=1
0.0261
‐0.0005
0.0006
0.0017
0.1797
‐0.0024
‐0.0037
‐0.0076
0.3756
‐0.0037
‐0.0058
‐0.0017
0.1504
0.0032
‐0.0004
0.0016
0.2438
0.0010
0.0025
0.0076
0.0192
0.0008
‐0.0025
‐0.0042
0.2736
0.0025
0.0033
0.0069
39
0.0241
‐0.0008
0.0017
0.0042
0.1699
‐0.0045
‐0.0073
‐0.0085
0.3186
‐0.0068
‐0.0175
‐0.0185
0.1749
0.0466
0.0438
‐0.0076
0.2148
‐0.0004
‐0.0009
‐0.0099
0.0312
‐0.0009
0.0008
‐0.0022
0.2941
0.0055
0.0099
0.0206
MERGER2=1
0.0239
0.0009
0.0034
0.0068
0.1673
‐0.0004
‐0.0056
‐0.0112
0.2901
‐0.0152
‐0.0237
‐0.0213
0.1235
‐0.0086
‐0.0096
‐0.0061
0.1667
‐0.0037
‐0.0122
‐0.0138
0.0382
‐0.0014
‐0.0004
‐0.0019
0.3018
0.0094
0.0165
0.0233
Table 5: Multinomial probit estimation results
(a) Baseline estimation
Multinomial probit estimation
Dependent variable: MTYPE1={0,1,2}
Marginal effect p>|z|
MTYPE1=1(MERGER1)
lnFIRMAGE
0.0025
lnEMP(t‐1)
0.0157 ***
lnASSETS(t‐1)
0.0333 ***
NBANK
0.0377 ***
RATING
0.0032 ***
HOKKAIDO
‐0.0490 TOHOKU
‐0.0789 ***
KANTO
0.3732 ***
KOSHINETSU
‐0.0916 ***
TOKAI
0.4740 ***
KINKI
0.3648 ***
CHUGOKU
0.0721 **
SHIKOKU
‐0.0047
KYUSHU
‐0.1825
Industry dummies
Yes
No. of Obs.
11107
Wald chi2 (40)
4017.62
p‐value
0.000
Log likelihood
‐6732.47
Std. err.
0.0087
0.0061
0.0054
0.0025
0.0008
0.0315
0.0290
0.0341
0.0289
0.0361
0.0384
0.0356
0.0359
0.0321
Marginal effect p>|z|
MTYPE1=2(MERGER2)
0.0057 0.0060 ***
0.0074 ***
0.0146 ***
0.0013 ***
‐0.0187 ***
‐0.0180 **
0.1269 ***
‐0.0320 ***
0.0729 ***
0.1471 ***
‐0.0092 ‐0.0184 **
‐0.0219 ***
Yes
Std. err.
0.0025
0.0015
0.0015
0.0015
0.0002
0.0061
0.0067
0.0279
0.0029
0.0264
0.0353
0.0083
0.0060
0.0049
(b) Estimation with four outcome values
Multinomial probit estimation
Dependent variable: MTYPE2={0,1,2,3}
Marginal effect p>|z|
Std. err.
Marginal effect p>|z|
MTYPE2=1(MERGER1_1)
MTYPE2=2(MERGER1_2)
lnFIRMAGE
0.0011
0.0061
0.0001 lnEMP(t‐1)
0.0046 0.0043
0.0110 ***
lnASSETS(t‐1)
0.0163 ***
0.0038
0.0160 ***
NBANK
0.0138 ***
0.0018
0.0229 ***
RATING
0.0010 **
0.0005
0.0020 ***
HOKKAIDO
‐0.0300
0.0202
‐0.0151
TOHOKU
‐0.0649 ***
0.0157
‐0.0035 KANTO
0.0848 ***
0.0245
0.2789 ***
KOSHINETSU
‐0.1029 ***
0.0097
0.0192 TOKAI
0.4674 ***
0.0426
‐0.0454 **
KINKI
0.2506 ***
0.0377
0.0895 ***
CHUGOKU
‐0.0267 0.0194
0.1015 ***
SHIKOKU
‐0.0421 **
0.0201
0.0219 KYUSHU
‐0.0717 ***
0.0135
0.0716 **
Industry dummies
Yes
Yes
No. of Obs.
11107
Wald chi2 (60)
4769.68
p‐value
0.000
Log likelihood
‐8222.79
Note: *,**,*** indicate significance at the 10%, 5%, and 1% level respectively.
40
Std. err.
0.0061
0.0041
0.0037
0.0018
0.0006
0.0245
0.0254
0.0351
0.0283
0.0183
0.0306
0.0344
0.0307
0.0313
Marginal effect p>|z|
MTYPE2=3(MERGER2)
0.0062 **
0.0063 ***
0.0078 ***
0.0154 ***
0.0013 ***
‐0.0197 ***
‐0.0191 ***
0.1361 ***
‐0.0337 ***
0.0886 ***
0.1624 ***
‐0.0093 ‐0.0194 ***
‐0.0229 ***
Yes
Std. err.
0.0027
0.0016
0.0016
0.0015
0.0003
0.0066
0.0072
0.0297
0.0031
0.0302
0.0381
0.0090
0.0065
0.0053
Table 6: Treatment effects estimation results
Baseline estimation
Treated: MTYPE1=2 (MERGER2)
Control: MTYPE1=0 (MERGER0)
(1)
(2)
DID
Std. err.
0.0010 0.0010
0.0020 *
0.0012
0.0040 ***
0.0014
Method: Caliper Matching Treated: MTYPE1=2 (MERGER2)
Control: MTYPE1=1 (MERGER1)
(3)
(4)
DID
Std. err.
0.0009
0.0007
0.0010
0.0009
0.0025 **
0.0012
Treated: MTYPE1=1 (MERGER1)
Control: MTYPE1=0 (MERGER0)
(5)
(6)
DID
Std. err.
‐0.0001
0.0007
0.0015 *
0.0008
0.0020 **
0.0009
RATE
t+1
t+2
t+3
lnRATE
t+1
t+2
t+3
0.0463
0.0710 *
0.1349 ***
0.0350
0.0426
0.0486
0.0282
0.0720 ***
0.0790 ***
0.0219
0.0254
0.0289
CASH
t+1
t+2
t+3
0.0076
0.0033
0.0072
0.0053
0.0060
0.0067
0.0024
0.0043
0.0010
0.0031
0.0038
0.0041
‐0.0006
‐0.0044
0.0008
0.0030
0.0034
0.0038
LOAN
t+1
t+2
t+3
‐0.0057
‐0.0078 ‐0.0051 0.0063
0.0082
0.0096
‐0.0051
‐0.0060 ‐0.0017 0.0039
0.0048
0.0058
‐0.0012
‐0.0034
‐0.0042
0.0036
0.0047
0.0056
SHORT
t+1
t+2
t+3
‐0.0074
‐0.0098
‐0.0037
0.0055
0.0070
0.0085
‐0.0031
‐0.0039
0.0000
0.0035
0.0040
0.0051
‐0.0016
‐0.0061
‐0.0056
0.0032
0.0041
0.0049
LONG
t+1
t+2
t+3
0.0014
0.0023
‐0.0009
0.0183
0.0204
0.0267
‐0.0007
‐0.0003
0.0002
0.0035
0.0044
0.0049
‐0.0001
0.0014
0.0006
0.0096
0.0108
0.0140
ROA
t+1
t+2
t+3
0.0000
0.0011
0.0041
0.0035
0.0041
0.0044
0.0017
0.0021
0.0025
‐0.0001
0.0021
0.0029
0.0020
0.0023
0.0025
0.0036
0.0041
0.0052
0.0018
‐0.0002
0.0074
0.0030
0.0040
0.0052
0.0001
0.0005 0.0002
CAP
t+1
0.0056 0.0054
‐0.0024
t+2
0.0096 0.0070
0.0011 t+3
0.0167 *
0.0092
0.0002 Note: *,**,*** indicate significance at the 10%, 5%, and 1% level respectively.
41
0.0088
0.0340
0.0686 **
0.0201
0.0240
0.0271
Table 7: Treatment effects estimation results
Estimation with four outcome values
Treated: MTYPE2=1 (MERGER1_1)
Control: MTYPE2=0 (MERGER0)
(1)
(2)
DID
Std. err.
0.0002
0.0008
0.0018 *
0.0011
0.0019 *
0.0011
RATE
t+1
t+2
t+3
lnRATE
t+1
t+2
t+3
CASH
t+1
t+2
t+3
‐0.0009
‐0.0032
0.0041
0.0039
0.0044
0.0049
LOAN
t+1
t+2
t+3
0.0015
‐0.0002 0.0026 SHORT
t+1
t+2
t+3
LONG
ROA
0.0083
0.0324 0.0536 0.0257
0.0308
0.0351
Method: Caliper Matching Treated: MTYPE2=2 (MERGER1_2)
Control: MTYPE2=0 (MERGER0)
(3)
(4)
DID
Std. err.
‐0.0005 0.0008
0.0012 0.0010
0.0022 **
0.0011
0.0034
0.0432 0.0879 ***
Treated: MTYPE2=1 (MERGER1_1)
Control: MTYPE2=2 (MERGER1_2)
(5)
(6)
DID
Std. err.
0.0002
0.0016
0.0003
0.0023
0.0008 0.0025
0.0231
0.0274
0.0306
0.0236
0.0167 0.0154 0.0456
0.0571
0.0611
‐0.0001
‐0.0021
0.0006
0.0035
0.0041
0.0045
0.0013 0.0013
0.0027
0.0071
0.0088
0.0093
0.0047
0.0062
0.0071
‐0.0068 *
‐0.0084 ‐0.0117 *
0.0041
0.0055
0.0064
0.0083
0.0116
0.0198
0.0088
0.0113
0.0135
‐0.0004
‐0.0060
‐0.0019
0.0043
0.0054
0.0065
‐0.0046
‐0.0051
‐0.0060
0.0036
0.0048
0.0057
0.0018
0.0000
0.0058 0.0081
0.0099
0.0114
t+1
t+2
t+3
0.0013
0.0035
0.0034
0.0116
0.0131
0.0170
‐0.0050
‐0.0066
‐0.0090
0.0036
0.0048
0.0056
0.0074
0.0114
0.0148 0.0082
0.0103
0.0107
t+1
t+2
t+3
‐0.0010
0.0009 0.0016
0.0025
0.0022
0.0025
0.0102
0.0115
0.0148
‐0.0031
‐0.0078 *
‐0.0100 *
0.0039
0.0046
0.0055
0.0035
0.0046
0.0061
‐0.0061
‐0.0150 ‐0.0130 0.0078
0.0093
0.0120
0.0002
0.0034
0.0049 *
CAP
t+1
‐0.0012
0.0038
0.0065 *
t+2
‐0.0033
0.0051
0.0053 t+3
0.0059 0.0066
0.0125 **
Note: *,**,*** indicate significance at the 10%, 5%, and 1% level respectively.
42
Table 8: Partial mean linear estimations for the treatment effect
(a) Baseline estimation
Baseline
Treated: MTYPE1=2 (MERGER2)
Treated: MTYPE1=2 (MERGER2)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=1 (MERGER1)
(1)
RATE
LnRATE
t+3
t+3
(2)
DID
0.0041 ***
0.1335 ***
Std. err.
0.0011
0.0310
(3)
Treated: MTYPE1=1 (MERGER1)
Control: MTYPE1=0 (MERGER0)
(4)
DID
0.0021 *
0.0627 **
(5)
Std. err.
0.0011
0.0270
(6)
DID
0.0020 ***
0.0708 ***
Std. err.
0.0007
0.0214
Note: *,**,*** indicate significance at the 10%, 5%, and 1% level respectively. Standard errors are calculated by the delta method utilizing the heteroskedasticity robust matrix estimator.
(b) Effect of bank branch consolidation
Consolidation of bank branches
ALL
Subsample
Treated: MTYPE1=2 (MERGER2)
Treated: MTYPE1=1 (MERGER1)
Treated: MTYPE1=2 (MERGER2)
Treated: MTYPE1=1 (MERGER1)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=0 (MERGER0)
DID
Std. err.
DID
Std. err.
DID
Std. err.
DID
Std. err.
RATE
t+3
0.0056 ***
0.0016
0.0036 ***
0.0013
0.0068 ***
0.0024
0.0028 *
0.0014
LnRATE
t+3
0.1693 ***
0.0405
0.1107 ***
0.0338
0.2093 **
0.0543
0.0915 **
0.0397
Notes: *,**,*** indicate significance at the 10%, 5%, and 1% level respectively. Standard errors are calculated by the delta method utilizing the heteroskedasticity robust matrix estimator.
In the s ubs a mpl e es ti ma ti on, treatment obs ervati ons a re l i mi ted to fi rms that s aw a cl os ure of thei r ba nk bra nch.
(c) Effect of firms' credit risk
RATE
LnRATE
t+3
t+3
RATE
LnRATE
t+3
t+3
RATING<median
ALL
Subsample
Treated: MTYPE1=2 (MERGER2)
Treated: MTYPE1=1 (MERGER1)
Treated: MTYPE1=2 (MERGER2)
Treated: MTYPE1=1 (MERGER1)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=0 (MERGER0)
DID
Std. err.
DID
Std. err.
DID
Std. err.
DID
Std. err.
0.0032 *
0.0018
0.0007 0.0009
0.0049 **
0.0019
0.0020 **
0.0009
0.0914 *
0.0499
0.0532 *
0.0311
0.1569 ***
0.0604
0.0977 ***
0.0369
SPREAD<0
ALL
Subsample
Treated: MTYPE1=2 (MERGER2)
Treated: MTYPE1=1 (MERGER1)
Treated: MTYPE1=2 (MERGER2)
Treated: MTYPE1=1 (MERGER1)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=0 (MERGER0)
Control: MTYPE1=0 (MERGER0)
DID
Std. err.
DID
Std. err.
DID
Std. err.
DID
Std. err.
0.0009
0.0016
0.0016 0.0011
0.0033 *
0.0020
0.0029 ***
0.0011
0.0953 **
0.0473
0.0462
0.0351
0.1260 **
0.0560
0.0598
0.0372
Notes: *,**,*** indicate significance at the 10%, 5%, and 1% level respectively. Standard errors are calculated by the delta method utilizing the heteroskedasticity robust matrix estimator.
In the s ubs a mpl e es ti ma ti on, the s ampl e i s l i mi ted to obs erva ti ons tha t s a ti s fy RATING<medi a n or SPREAD<0.
43
0
0
10
10
20
20
30
30
40
40
Figure 1: Distributions of borrowing costs by subsample
0
.02
.04
.06
RATE(MERGER2, Truncated)
.1
0
RATE(Year 2008)
30
20
10
0
0
.02
.04
.06
RATE(MERGER1, Truncated)
RATE(Year 2005)
.08
.02
.04
.06
RATE(MERGER0, Truncated)
RATE(Year 2005)
40
RATE(Year 2005)
.08
.1
RATE(Year 2008)
44
.08
RATE(Year 2008)
.1
Figure 2: Treatment effect depending on the number of firm‐bank relationships
Treatment: MTYPE1=1 (MERGER1) and Control: MTYPE1=0 (MERGER0)
140
difference-in-differences (basis points)
120
100
80
60
40
20
0
-20
-40
2
3
4
5
6
7
8
9
10
number of banks
Treatment: MTYPE1=2 (MERGER2) and Control: MTYPE1=0 (MERGER0)
140
difference-in-difference (basis points)
120
100
80
60
40
20
0
-20
-40
2
3
4
5
6
7
8
9
10
number of banks
Note: The dotted lines represent the 95% confidence interval. Standard errors are estimated
by bootstrapping (non‐parametric bootstrap) with 1,000 replications.
45